Railway freight train state evaluation method based on time sequence and spectral analysis

Abstract

The invention discloses a railway freight train state evaluation method based on a time sequence and spectral analysis, and relates to the technical field of railway transportation safety monitoring, and the method comprises the steps: collecting time sequence data and image data during the operation of a train, and carrying out the preprocessing of the time sequence data and the image data; splicing and registering the preprocessed image data, generating a full-scene image and extracting dynamic visual features; performing multi-scale time sequence spectrum conjoint analysis on the preprocessed time sequence data, and performing multi-scale fusion in combination with visual features to generate a multi-dimensional feature matrix; constructing a health state baseline library, comparing the multi-dimensional feature matrix with the health state baseline library, and identifying abnormal features; and constructing a train state evaluation model, evaluating the train operation state according to the abnormal characteristics, and generating a state evaluation report according to an evaluation result. According to the method, the comprehensiveness and accuracy of train state monitoring are remarkably improved through multi-scale time sequence spectrum conjoint analysis and visual feature fusion.

Description

Technical Field

[0001] The invention relates to the technical field of railway transportation safety monitoring, in particular to a railway freight train state assessment method based on time series and spectrum analysis. Background Art

[0002] With the rapid development of railway transportation technology, the increase in train speed and load capacity has made condition monitoring and evaluation technology an important means to ensure the safety of train operation. At present, train condition monitoring mostly relies on data acquisition and analysis methods of a single signal source, such as vibration monitoring, temperature monitoring or strain monitoring. These methods identify abnormal conditions in train operation by analyzing the spectrum or time domain of a single type of sensor data. However, this single signal source monitoring method is difficult to fully reflect the complexity of the train operation status, especially when multiple fault factors interact, its diagnostic ability has obvious limitations. In addition, with the advancement of image processing technology, visual means of train condition monitoring has gradually received attention, such as using cameras to detect cracks or wear on the appearance of the train. However, existing image monitoring technologies mostly target single points or single components, lack the ability to integrate monitoring of the entire train, and the accuracy of image stitching and registration technology is not enough to meet the needs of dynamic monitoring when the train is running at high speed.

[0003] The shortcomings of existing technologies are mainly reflected in the following two points: first, it is difficult to achieve the fusion of multimodal data based on the analysis of a single signal source, and it is impossible to accurately and comprehensively evaluate the overall status of the train; second, image monitoring technology fails to achieve efficient splicing and integration of multi-camera images, and it is difficult to provide visualization support for the status of the train in all scenarios. Therefore, a multimodal evaluation method that can integrate time series, spectrum analysis and visual features is needed to improve the comprehensiveness, accuracy and real-time performance of train status monitoring. Summary of the invention

[0004] In view of the above existing problems, the present invention is proposed.

[0005] Therefore, the present invention provides a railway freight train status assessment method based on time series and spectrum analysis to solve the problems of insufficient multimodal data fusion and low efficiency of image monitoring splicing and integration in the prior art.

[0006] In order to solve the above technical problems, the present invention provides the following technical solutions:

[0007] In a first aspect, the present invention provides a railway freight train status assessment method based on time series and spectrum analysis, which includes collecting time series data and image data during train operation, and preprocessing the time series data and image data; splicing and aligning the preprocessed image data to generate a full-scene image and extract dynamic visual features; performing multi-scale time series spectrum joint analysis on the preprocessed time series data, and performing multi-scale fusion in combination with visual features to generate a multidimensional feature matrix; constructing a health status baseline library, comparing the multidimensional feature matrix with the health status baseline library, and identifying abnormal features; constructing a train status assessment model, evaluating the train operation status according to the abnormal features, and generating a status assessment report according to the assessment results.

[0008] As a preferred solution of the railway freight train status assessment method based on time series and spectrum analysis of the present invention, wherein: the time series data includes vibration data, temperature data and load data during train operation;

[0009] The image data includes a crack image, a wear image, and an oil stain image of the train appearance.

[0010] As a preferred solution of the railway freight train status assessment method based on time series and spectrum analysis of the present invention, the specific steps of preprocessing the time series data and image data are as follows:

[0011] A bandpass filter is used to denoise the vibration data;

[0012] The moving average method is used to smooth the temperature data;

[0013] Use wavelet transform to remove random noise in drinking data;

[0014] Image data is enhanced through histogram equalization;

[0015] Image data is normalized by adjusting the image size and resolution.

[0016] As a preferred solution of the railway freight train status assessment method based on time series and spectrum analysis of the present invention, the specific steps of splicing and registering the pre-processed image data, generating a full scene image and extracting visual features are as follows:

[0017] Use SIFT to identify feature points from preprocessed image data and extract feature descriptors from each feature point. Based on the feature descriptors, use the FLANN feature matching method to match feature point pairs between different image data.

[0018] The RANSAC algorithm is used to randomly select the minimum number of feature point pairs from the matched feature point pairs, and a geometric transformation model is constructed based on the selected feature point pairs;

[0019] According to the geometric transformation model, the position and angle of the image data are adjusted, the image data is spatially registered, and the registered image data is fused through a multi-resolution pyramid to form a full-scene image;

[0020] Apply time series segmentation to the full scene image and use temporal optical flow to track the dynamic changes of the full scene image over time;

[0021] Based on the dynamic changes of the full scene image over time, the dynamic visual features of the train are extracted from the full scene image using edge detection operators and CNN.

[0022] As a preferred solution of the railway freight train status assessment method based on time series and spectrum analysis described in the present invention, wherein: the pre-processed time series data is subjected to multi-scale time series spectrum joint analysis, and multi-scale fusion is performed in combination with visual features to generate a multi-dimensional feature matrix. The specific steps are as follows:

[0023] The vibration data is converted into time-frequency representation by short-time Fourier transform, and the dynamic spectrum features within a fixed time window are extracted based on the time-frequency representation;

[0024] LSTM is used to capture the temporal dependencies between temperature data and load data, and to extract low-frequency trend features within a long time window.

[0025] Through the multi-scale feature fusion method, the spectrum features, low-frequency trend features and visual features are fused to generate a multi-dimensional feature matrix, and the timestamp is added to the multi-dimensional feature matrix. The expression is:

[0026]

[0027] Among them, F is a multidimensional feature matrix, D N is the dynamic spectrum feature vector of the Nth sample point, L N is the low-frequency trend feature vector of the Nth sample point, V N is the dynamic visual feature vector of the Nth sample point, D 1 is the dynamic spectrum feature vector of the first sample point, D 2 is the dynamic spectrum feature vector of the second sample point, L 1 is the low-frequency trend feature vector of the second sample point, L 2 is the low-frequency trend feature vector of the second sample point, V 1 is the dynamic visual feature vector of the first sample point, V 2is the dynamic visual feature vector of the second sample point, N is the total number of samples, t 1 is the timestamp of the first sample point, t 2 is the timestamp of the second sample point, t N is the timestamp of the Nth sample point, and T represents the transposition symbol.

[0028] As a preferred solution of the railway freight train status assessment method based on time series and spectrum analysis of the present invention, wherein: the health status baseline library is constructed, the multidimensional feature matrix is ​​compared with the health status baseline library, and abnormal features are identified. The specific steps are as follows:

[0029] Collect historical train operation data and conduct statistical analysis on the historical train operation data to identify the mean, variance and statistics of historical dynamic visual features, dynamic spectrum features and low-frequency trend features, and build a health status baseline library based on the identification results;

[0030] The multidimensional feature matrix is ​​compared with the health status baseline library, and the outlier index of each feature in the multidimensional feature matrix is ​​predicted based on the interaction between the features. The expression is:

[0031]

[0032] Among them, A(F) is the outlier index of each sample point feature in the multidimensional feature matrix, D k is the dynamic spectrum feature vector in the kth time window, μ D is the mean vector of the dynamic spectrum feature, σ D is the standard deviation vector of the dynamic spectrum feature, L k is the low-frequency trend feature vector in the kth time window, μ L is the mean vector of the low-frequency trend feature, σ L is the standard deviation vector of low-frequency trend features, V k is the dynamic visual feature vector in the kth time window, μ V is the mean vector of dynamic visual features, σ V is the standard deviation vector of dynamic visual features, ω(t k ) is the time weighting factor, t k is the timestamp corresponding to the kth time window, λ is the weight coefficient of the interaction term, I(D k ,V k ) is the Pearson correlation coefficient between the dynamic spectrum feature vector and the dynamic visual feature vector, k is the index variable of the time window;

[0033] Based on historical data, anomaly detection threshold E is defined. When A(F) is greater than E, the sample point features in the current time window are considered to be abnormal features.

[0034] As a preferred solution of the railway freight train state assessment method based on time series and spectrum analysis of the present invention, the Pearson correlation coefficient between the dynamic spectrum feature vector and the dynamic visual feature vector is expressed as:

[0035]

[0036] Among them, D k,i is the i-th element of the dynamic spectrum feature vector in the k-th time window, V k,i is the i-th element of the dynamic visual feature vector in the k-th time window, m represents the number of elements in each feature vector, and i represents the index variable of the element in each feature vector.

[0037] As a preferred solution of the railway freight train status assessment method based on time series and spectrum analysis of the present invention, wherein: the train status assessment model is constructed, the train running status is assessed according to abnormal characteristics, and a status assessment report is generated according to the assessment results. The specific steps are as follows:

[0038] Using LSTM as the basis, a multi-layer LSTM unit is set up. Each unit is responsible for capturing the local time dependency within a time window. The LSTM layer receives the abnormal features as input and outputs the time dependency representation of the abnormal features.

[0039] The Transformer structure is combined with the LSTM layer, and the dynamic change characteristics in the time-dependent representation of abnormal features are processed through a feedforward neural network;

[0040] The LSTM layer and the Transformer structure are integrated through the fully connected layer to form a train status assessment model;

[0041] The abnormal features are input into the train status assessment model to predict the train operation status score. The expression is:

[0042]

[0043] Where S is the train running status score, Y is the score calibration coefficient, ΔB(t) is the abnormal feature vector at time point t, H(t) represents the dynamic weight of the train running status score at time point t, and t 0 Indicates the starting time of the time window, t e represents the end time of the time window, ξ represents the weight coefficient of the abnormal feature vector modulus, ΔB k The abnormal feature vector in the kth time window, R k is the risk coefficient in the kth time window, λ is the weight coefficient of the train operation cost factor, C kis the train operation cost factor in the kth time window;

[0044] Based on the statistical analysis of historical data, the health status benchmark value Q is defined;

[0045] When S ≥ Q, the train is considered to be in good running condition;

[0046] When S<Q, the train running status is considered abnormal;

[0047] Based on the assessment results, a status assessment report is generated.

[0048] In a second aspect, the present invention provides a computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: when the computer program is executed by the processor, any step of the railway freight train status assessment method based on time series and spectrum analysis as described in the first aspect of the present invention is implemented.

[0049] In a third aspect, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the railway freight train status assessment method based on time series and spectral analysis as described in the first aspect of the present invention.

[0050] The beneficial effects of the present invention are as follows: through the joint analysis of multi-scale time series spectrum and the fusion of visual features, the comprehensiveness and accuracy of train status monitoring are significantly improved, ensuring that abnormal situations will not be missed, and improving the reliability of fault prediction. At the same time, by efficiently splicing and registering the pre-processed image data, generating full-scene images and extracting dynamic visual features, the present invention provides intuitive visualization support for the appearance of the entire train, enhancing the real-time performance and maintenance efficiency of the status monitoring system, thereby greatly improving the accuracy and response speed of railway freight train status assessment. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] In order to more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings required for use in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other accompanying drawings can be obtained based on these accompanying drawings without paying creative work.

[0052] Figure 1 This is a flow chart of the railway freight train status assessment method based on time series and spectrum analysis in Example 1.

[0053] Figure 2 This is a flowchart for preprocessing time series data and image data in Example 1. DETAILED DESCRIPTION

[0054] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the specific implementation methods of the present invention are described in detail below in conjunction with the accompanying drawings.

[0055] In the following description, many specific details are set forth to facilitate a full understanding of the present invention, but the present invention may also be implemented in other ways different from those described herein, and those skilled in the art may make similar generalizations without violating the connotation of the present invention. Therefore, the present invention is not limited to the specific embodiments disclosed below.

[0056] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The term "in one embodiment" that appears in different places in this specification does not necessarily refer to the same embodiment, nor does it refer to a separate or selective embodiment that is mutually exclusive with other embodiments.

[0057] Example 1, reference Figure 1 and Figure 2 , which is the first embodiment of the present invention, provides a railway freight train status assessment method based on time series and spectrum analysis, comprising the following steps:

[0058] S1: Collect time series data and image data during train operation, and pre-process the time series data and image data.

[0059] S1.1: Time series data include vibration data, temperature data and load data during train operation;

[0060] Furthermore, the time series data are collected in real time by sensors installed on key parts of the train (such as axles, bearings, and bogies);

[0061] S1.2: The image data includes crack images, wear images, and oil stain images of the train appearance.

[0062] Furthermore, the image data is captured by a portal frame acquisition system, covering key areas of the train appearance (such as wheels, rail contact surfaces, and car connections);

[0063] S1.3: Use a bandpass filter to denoise the vibration data;

[0064] For example, by setting the frequency range of the bandpass filter to 20 Hz to 200 Hz, high-frequency noise and low-frequency interference in the vibration data can be effectively removed to ensure the purity of the signal.

[0065] S1.4: Use the moving average method to smooth the temperature data;

[0066] For example, using a moving average of the data points with a window size of 5 smooths out short-term fluctuations in the temperature data, making the trend more apparent.

[0067] S1.5: Use wavelet transform to remove random noise in drinking data;

[0068] For example, the Daubechies wavelet transform is applied to decompose the reheat data, retaining the main signal components and removing high-frequency random noise, thus improving the data quality.

[0069] S1.6: Enhance the image data by histogram equalization;

[0070] For example, applying histogram equalization to a crack image enhances the contrast of the image, making subtle cracks more clearly visible.

[0071] S1.7: Normalize the image data by adjusting the image size and resolution.

[0072] For example, all collected wear images are uniformly adjusted to 512x512 pixels and the resolution is set to 300DPI to ensure the consistency and comparability of the image data.

[0073] S2: Stitch and align the preprocessed image data to generate a full-scene image and extract dynamic visual features.

[0074] S2.1: Use SIFT to identify feature points from the preprocessed image data and extract feature descriptors from each feature point. Based on the feature descriptors, use the FLANN feature matching method to match feature point pairs between different image data.

[0075] The specific process is as follows: first use the SIFT (Scale-Invariant Feature Transform) algorithm to detect and identify stable feature points, and extract detailed feature descriptors from each feature point. These descriptors can capture the local texture information around the feature point. Then, based on the extracted feature descriptors, the FLANN (Fast Library for Approximate Nearest Neighbors) feature matching method is used to efficiently find and match feature point pairs between different image data. For example, assuming there are two images of the appearance of a train from different cameras, the SIFT algorithm will find multiple significant feature points in each image and generate corresponding 128-dimensional feature descriptors. Then, the FLANN algorithm will quickly perform an approximate nearest neighbor search between the two sets of feature descriptors to find the most similar feature point pairs. In this way, even if the image has scale, rotation or illumination changes, the feature points can be accurately matched, thereby providing a reliable correspondence for subsequent image stitching and spatial registration, ensuring that the generated full-scene image has high accuracy and consistency.

[0076] S2.2: Use the RANSAC algorithm to randomly select the minimum number of feature point pairs from the matched feature point pairs, and build a geometric transformation model based on the selected feature point pairs;

[0077] The specific process is as follows: the RANSAC algorithm randomly selects a set of feature point pairs each time, tries to build a preliminary geometric transformation model, and then calculates the degree of conformity of all other feature point pairs with the model. Feature point pairs that conform to the geometric transformation model are considered inliers, and those that do not conform are outliers. After multiple iterations, the model with the most inliers is retained as the final geometric transformation model, ensuring the robustness and accuracy of the geometric transformation model. For example, in the process of train image stitching, the RANSAC algorithm can effectively handle feature point matching errors caused by differences in camera perspectives and environmental changes, thereby generating accurate spatial registration results.

[0078] S2.3: According to the geometric transformation model, the position and angle of the image data are adjusted, the image data is spatially registered, and the registered image data is fused through a multi-resolution pyramid to form a full-scene image;

[0079] The specific process is: apply the geometric transformation model to transform each image to ensure that all images are aligned in a unified coordinate system. Then, the registered image data is stitched through the multi-resolution pyramid fusion technology, which first roughly aligns the images at the low-resolution level, gradually transitions to the high-resolution level for fine adjustment, and finally generates a seamless and high-quality full-scene image. For example, in train appearance monitoring, this process can perfectly stitch together images from different perspectives from multiple cameras to provide a complete view of the train appearance, ensuring that no details are missed.

[0080] S2.4: Apply time series segmentation to the full scene image and use temporal optical flow to track the dynamic changes of the full scene image over time;

[0081] The specific process is as follows: the temporal optical flow algorithm can accurately estimate the movement direction and speed of each pixel in the image, generate a continuous time series dynamic graph, and thus achieve real-time monitoring of the train status and early warning of abnormal changes. For example, in the status assessment of railway freight trains, this method can effectively detect small changes in the surface features of the train.

[0082] S2.5: Based on the dynamic changes of the full scene image over time, the dynamic visual features of the train are extracted from the full scene image using edge detection operators and CNN.

[0083] The specific process is as follows: first, edge detection operators (such as Canny or Sobel operators) are applied to process the full-scene image to identify and enhance significant edges and contours in the image, and to capture changes in the train's appearance features. Then, a convolutional neural network (CNN) is used to perform in-depth analysis of the full-scene image after edge detection processing to extract more representative dynamic visual features, such as crack extension, increased wear, and oil diffusion. Specifically, the edge detection operator can highlight important structural information in the full-scene image, while the convolutional neural network automatically learns and extracts complex patterns and features through the operation of multiple layers of convolutional layers and pooling layers, and finally generates a high-dimensional feature vector for a comprehensive description of the changes in the train's appearance over time. This process ensures that the dynamic visual features extracted from the full-scene image are both accurate and representative, providing key data support for subsequent status assessments.

[0084] S3: Perform multi-scale time series spectrum joint analysis on the preprocessed time series data, and combine it with visual features for multi-scale fusion to generate a multi-dimensional feature matrix.

[0085] S3.1: Convert the vibration data into time-frequency representation by short-time Fourier transform, and extract the dynamic spectrum features within a fixed time window based on the time-frequency representation;

[0086] The specific process is as follows: STFT applies Fourier transform to each time window to generate a spectrum within the window, from which the intensity and distribution of specific frequency components can be identified. These dynamic spectrum features reflect the frequency composition and change trend of vibration data at different time points, providing an important basis for subsequent multi-scale time series spectrum joint analysis. This process ensures the accurate capture and analysis of potential abnormal patterns in vibration data.

[0087] S3.2: Use LSTM to capture the temporal dependencies between temperature and load data and extract low-frequency trend features within a long time window;

[0088] The specific process is as follows: the LSTM network receives the preprocessed temperature data and load data as input, and updates its internal state at each time step, thereby learning and capturing the hidden time-dependent patterns in these data. This feature of LSTM enables it to effectively process long-time series data and identify complex dependencies across multiple time steps. On this basis, low-frequency trend features within a long time window are further extracted from the LSTM output. These features reflect the slow-changing trends of temperature and load data on a longer time scale, such as a gradually increasing or decreasing pattern. This process ensures the accurate capture of the long-term evolution characteristics of the train's running status, and provides important time dimension information for the subsequent joint analysis of multi-scale time series spectra.

[0089] S3.3: Through the multi-scale feature fusion method, the spectrum features, low-frequency trend features and visual features are fused to generate a multi-dimensional feature matrix, and the timestamp is added to the multi-dimensional feature matrix. The expression is:

[0090]

[0091] Among them, F is a multidimensional feature matrix, D N is the dynamic spectrum feature vector of the Nth sample point, L N is the low-frequency trend feature vector of the Nth sample point, V N is the dynamic visual feature vector of the Nth sample point, D 1 is the dynamic spectrum feature vector of the first sample point, D 2 is the dynamic spectrum feature vector of the second sample point, L 1 is the low-frequency trend feature vector of the second sample point, L 2 is the low-frequency trend feature vector of the second sample point, V 1 is the dynamic visual feature vector of the first sample point, V 2 is the dynamic visual feature vector of the second sample point, N is the total number of samples, t 1 is the timestamp of the first sample point, t 2 is the timestamp of the second sample point, tN is the timestamp of the Nth sample point, and T represents the transposition symbol.

[0092] It should be noted that the dynamic spectrum feature vector, low-frequency trend feature vector and dynamic visual feature vector of each sample point are first integrated, and the corresponding timestamp is added to each row to form a multidimensional feature matrix. Each row of the matrix represents all the features of a sample point and its time information, thus comprehensively describing the state changes of the train at different time points. This process ensures the effective fusion of multimodal data and provides comprehensive and detailed data support for subsequent state evaluation.

[0093] S4: Build a health status baseline library, compare the multidimensional feature matrix with the health status baseline library, and identify abnormal features.

[0094] S4.1: Collect historical train operation data and perform statistical analysis on the historical train operation data to identify the mean, variance and statistics of historical dynamic visual features, dynamic spectrum features and low-frequency trend features, and build a health status baseline library based on the identification results;

[0095] The specific process is as follows: First, a large amount of historical train operation data is collected, which covers the normal operation in different time periods. Then, a comprehensive statistical analysis is performed on these data to calculate the mean, variance and other related statistics of each feature (dynamic visual features, dynamic spectrum features and low-frequency trend features). For example, for dynamic spectrum features, the average frequency distribution and its fluctuation range in different time windows are calculated; for low-frequency trend features, the long-term change trend and stability of temperature and load data over time are evaluated; for dynamic visual features, the frequency and severity of features such as cracks, wear and oil stains in the image are analyzed. Based on the results of these statistical analyses, a health status baseline library is established, which records the feature distribution of the train under normal operation and provides a reference standard for subsequent status assessment. This process ensures the accuracy and reliability of the health status baseline library, which can effectively support anomaly detection and status assessment.

[0096] S4.2: Compare the multidimensional feature matrix with the health status baseline library, and combine the interaction between the features to predict the outlier index of each feature in the multidimensional feature matrix. The expression is:

[0097]

[0098] Among them, A(F) is the outlier index of each sample point feature in the multidimensional feature matrix, D k is the dynamic spectrum feature vector in the kth time window, μ D is the mean vector of the dynamic spectrum feature, σ D is the standard deviation vector of the dynamic spectrum feature, Lk is the low-frequency trend feature vector in the kth time window, μ L is the mean vector of the low-frequency trend feature, σ L is the standard deviation vector of low-frequency trend features, V k is the dynamic visual feature vector in the kth time window, μ V is the mean vector of dynamic visual features, σ V is the standard deviation vector of dynamic visual features, ω(t k ) is the time weighting factor, t k is the timestamp corresponding to the kth time window, λ is the weight coefficient of the interaction term, I(D k ,V k ) is the Pearson correlation coefficient between the dynamic spectrum feature vector and the dynamic visual feature vector, k is the index variable of the time window;

[0099] It should be noted that the dynamic spectrum feature vector, low-frequency trend feature vector and dynamic visual feature vector of each sample point in the multidimensional feature matrix are first compared with the corresponding mean vector and standard deviation vector in the health status baseline library, and the deviation of these feature vectors from the historical normal state is calculated. Then, the time weighting factor is comprehensively considered to evaluate the importance of different time points. In addition, the interaction weight coefficient and Pearson correlation coefficient are introduced to quantify the interaction between the dynamic spectrum feature vector and the dynamic visual feature vector. Finally, by combining the above deviation values ​​and interaction relationships, the abnormal index of the features of each sample point in the multidimensional feature matrix is ​​calculated, thereby realizing an accurate assessment of the train operation status.

[0100] S4.3: Based on historical data, define anomaly detection threshold E. When A(F) is greater than E, the sample point features in the current time window are considered to be abnormal features.

[0101] Furthermore, a large amount of historical operation data is collected and analyzed to identify the distribution of various features under normal conditions. Then, based on these statistical results, a threshold is set that can distinguish between normal and abnormal conditions. When the abnormal index of each sample point feature in the calculated multidimensional feature matrix exceeds this threshold, the sample point feature in the current time window is considered to be an abnormal feature. This process ensures the accuracy and reliability of anomaly detection and can promptly detect potential problems in train operation.

[0102] S4.4: Pearson correlation coefficient between dynamic spectrum feature vector and dynamic visual feature vector, expressed as:

[0103]

[0104] Among them, D k,i is the i-th element of the dynamic spectrum feature vector in the k-th time window, Vk,i is the i-th element of the dynamic visual feature vector in the k-th time window, m represents the number of elements in each feature vector, and i represents the index variable of the element in each feature vector.

[0105] It should be noted that for each time window k, the deviations of the elements at the corresponding positions in the dynamic spectrum feature vector and the dynamic visual feature vector are calculated respectively. Then, these deviations are multiplied and summed to obtain the numerator. Next, the sum of the squares of the deviations of each element in the dynamic spectrum feature vector and the dynamic visual feature vector are calculated respectively, and the square root is taken to obtain the denominator. Finally, the Pearson correlation coefficient between the dynamic spectrum feature vector and the dynamic visual feature vector is calculated by dividing the numerator by the denominator. This process quantifies the degree of linear correlation between the two feature vectors and provides important interactive relationship information for evaluating the train status.

[0106] S5: Build a train status assessment model, assess the train operation status according to abnormal characteristics, and generate a status assessment report based on the assessment results.

[0107] S5.1: Using LSTM as the basis, a multi-layer LSTM unit is set up, each unit is responsible for capturing the local temporal dependency within a time window, the LSTM layer receives the abnormal features as input, and outputs the temporal dependency representation of the abnormal features;

[0108] The specific process is as follows: the LSTM layer receives abnormal features as input, and processes these features step by step through multiple layers of LSTM units, each of which focuses on extracting local time dependencies within a specific time window. After multiple layers of processing, the LSTM layer outputs a time-dependent representation of the abnormal features, which can capture the evolution pattern and potential correlation of the abnormal features in the time series. This process ensures in-depth analysis of the abnormal features and provides strong time dimension information support for subsequent state evaluation.

[0109] S5.2: Combine the Transformer structure based on the LSTM layer and use a feedforward neural network to process the dynamic changes in the time-dependent representation of abnormal features;

[0110] The specific process is as follows: After the LSTM layer captures the time-dependent representation of abnormal features, the Transformer structure is further introduced, and its multi-head self-attention mechanism is used to extract global features of time series data to enhance the understanding of long-term dependencies. Subsequently, these time-dependent representations are processed through a feedforward neural network to capture the dynamic change characteristics, such as the evolution pattern and potential trend of abnormal features over time. This process ensures that the abnormal features are not only analyzed in terms of local time dependency, but also have a comprehensive understanding of global features and dynamic characteristics, providing more comprehensive and accurate data support for subsequent state assessment.

[0111] S5.3: The LSTM layer is integrated with the Transformer structure through a fully connected layer to form a train status assessment model;

[0112] The specific process is: the time dependency representation of the LSTM layer and the global feature representation of the Transformer structure are deeply integrated through the fully connected layer to integrate the advantages of both. The fully connected layer maps these fused feature representations to a higher level of abstract representation through linear and nonlinear transformations, and finally forms a complete train status assessment model. This process ensures that the model can capture both local time dependency and global features, providing more comprehensive and accurate status assessment results.

[0113] S5.4: Input the abnormal features into the train status assessment model to predict the train operation status score, the expression is:

[0114]

[0115] Where S is the train running status score, Y is the score calibration coefficient, ΔB(t) is the abnormal feature vector at time point t, H(t) represents the dynamic weight of the train running status score at time point t, and t 0 Indicates the starting time of the time window, t e represents the end time of the time window, ξ represents the weight coefficient of the abnormal feature vector modulus, ΔB k The abnormal feature vector in the kth time window, R k is the risk coefficient in the kth time window, λ is the weight coefficient of the train operation cost factor, C k is the train operation cost factor in the kth time window;

[0116] It should be noted that by comprehensively analyzing the changes of these abnormal features over time and their dynamic weights, a score reflecting the overall operation status of the train is calculated. In the scoring process, the weight of the modulus of the abnormal feature vector, the risk coefficient in each time window, and the influence of the train operation cost factor are considered. The final generated train operation status score not only reflects the comprehensive health level of the current state, but also considers the importance of abnormal features in different time windows and their potential risks, providing a comprehensive and accurate result for the train status assessment.

[0117] S5.5: Based on the statistical analysis of historical data, define the health status benchmark value Q;

[0118] When S ≥ Q, the train is considered to be in good running condition;

[0119] When S<Q, the train running status is considered abnormal;

[0120] It should be noted that this process ensures the objectivity and consistency of the status assessment through quantitative assessment standards, and reduces the subjective errors of human judgment. At the same time, the accuracy and reliability of the reference value Q are improved by combining historical data and expert experience, so that potential problems in train operation can be identified in a timely and accurate manner, ensuring the safe operation and maintenance efficiency of the train.

[0121] S5.6: Generate a status assessment report based on the assessment results.

[0122] It should be noted that the status assessment report includes train operation status score, abnormal characteristic analysis, potential risk warning and maintenance recommendations.

[0123] This embodiment also provides a computer device, which is suitable for the case of a railway freight train status assessment method based on time series and spectrum analysis, including: a memory and a processor; the memory is used to store computer executable instructions, and the processor is used to execute computer executable instructions to implement the railway freight train status assessment method based on time series and spectrum analysis proposed in the above embodiment.

[0124] The computer device may be a terminal, and the computer device includes a processor, a memory, a communication interface, a display screen and an input device connected via a system bus. The processor of the computer device is used to provide computing and control capabilities. The memory of the computer device includes a non-volatile storage medium and an internal memory. The non-volatile storage medium stores an operating system and a computer program. The internal memory provides an environment for the operation of the operating system and the computer program in the non-volatile storage medium. The communication interface of the computer device is used to communicate with an external terminal in a wired or wireless manner, and the wireless manner can be achieved through WIFI, an operator network, NFC (near field communication) or other technologies. The display screen of the computer device may be a liquid crystal display screen or an electronic ink display screen, and the input device of the computer device may be a touch layer covering the display screen, or a key, trackball or touchpad provided on the housing of the computer device, or an external keyboard, touchpad or mouse, etc.

[0125] This embodiment also provides a storage medium on which a computer program is stored. When the program is executed by a processor, the railway freight train status assessment method based on time series and spectrum analysis proposed in the above embodiment is implemented; the storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (Static Random Access Memory, referred to as SRAM), electrically erasable programmable read-only memory (Electrically Erasable Programmable Read-Only Memory, referred to as EEPROM), erasable programmable read-only memory (Erasable Programmable Read Only Memory, referred to as EPROM), programmable read-only memory (Programmable Red-Only Memory, referred to as PROM), read-only memory (Read-Only Memory, referred to as ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0126] In summary, the present invention significantly improves the comprehensiveness and accuracy of train status monitoring through: joint analysis of multi-scale time series spectrum and fusion of visual features, ensures that abnormal situations will not be missed, and improves the reliability of fault prediction. At the same time, by efficiently splicing and aligning the pre-processed image data, generating full-scene images and extracting dynamic visual features, the present invention provides intuitive visualization support for the appearance of the entire train, enhances the real-time performance and maintenance efficiency of the status monitoring system, and thus greatly improves the accuracy and response speed of railway freight train status assessment.

[0127] Example 2, referring to Table 1, is the second example of the present invention. To further verify the technical solution of the present invention, experimental simulation data of a railway freight train state assessment method based on time series and spectrum analysis are provided.

[0128] In order to verify the effectiveness and superiority of the invented method, a busy freight railway line was selected as the experimental site, and three running trains (Trains A, B, and C) were selected for a one-month monitoring experiment. During the experiment, in order to ensure the comprehensiveness and accuracy of data collection, high-precision sensors were installed on key parts of each train (such as axles, bearings, and bogies) to collect vibration data, temperature data, and load data in real time. At the same time, when the train passes, the door frame acquisition system automatically captures images of the train's appearance, covering key areas such as wheels, track contact surfaces, and car connections.

[0129] First, in order to verify the effectiveness of the invented method, a busy freight railway line was selected and 10 trains were selected for monitoring for a month. High-precision sensors were installed at key parts of the trains to collect vibration, temperature and load data in real time; at the same time, a door frame system was used to automatically capture images of the train's exterior to ensure that the data source was comprehensive and accurate.

[0130] Secondly, the collected data was preprocessed: the vibration data was denoised using a bandpass filter, the temperature data was smoothed using the moving average method, and the reheat data was denoised using wavelet transform; the image data was enhanced in contrast by histogram equalization and adjusted to a uniform size. Then, the SIFT algorithm was used to detect feature points, FLANN was used to match feature point pairs, and RANSAC was used to build a geometric transformation model to splice the images into a full scene image and extract dynamic visual features.

[0131] Next, the vibration data is converted into a time-frequency representation, and the time dependency of the temperature and load data is captured by LSTM to generate a multidimensional feature matrix. Based on the historical data analysis results, a health status baseline library is constructed and compared with the multidimensional feature matrix to identify abnormal features. This improves the accuracy of data analysis and enhances the early warning capability of abnormal situations.

[0132] Finally, the train running status is evaluated according to the abnormal characteristics and a status evaluation report is generated. The entire experiment strictly follows the above steps to ensure the authenticity and reliability of the data, demonstrating the significant advantages of the present invention in improving train safety and maintenance efficiency.

[0133] The existing technology uses traditional sensor networks to collect vibration, temperature and load data. The data processing of these sensors mainly relies on simple filtering and statistical analysis methods. For image data, the existing technology uses a conventional visual inspection system, which can capture the appearance of the train, but the image enhancement and standardization processing are relatively basic and cannot significantly improve the recognition accuracy of subtle defects (such as cracks and wear). In addition, the existing technology mainly relies on fixed threshold judgments and limited machine learning models for feature extraction.

[0134] The details are shown in Table 1 below:

[0135] Table 1 Comparison of train status monitoring data

[0136]

[0137] Through the analysis of the data in the above table, it can be clearly seen that the method of the present invention is significantly better than the existing technology in various performance indicators of train status monitoring. For example, in terms of vibration quality index (VQI), the new method reached an average of 9.03, while the existing technology was only 7.5; the temperature stability index (TSI) was also improved from 3.8 to 4.6, indicating that the present invention can more accurately process and analyze time series data, thereby providing more stable temperature monitoring.

[0138] The railway freight train status assessment method based on time series and spectrum analysis of the present invention can not only significantly improve the accuracy and efficiency of train operation data processing, but also achieve real-time, high-precision monitoring of train status and abnormal warning. This reflects the significant contribution of the present invention to improving railway transportation safety, reliability and maintenance efficiency.

[0139] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention rather than to limit it. Although the present invention has been described in detail with reference to the preferred embodiments, those skilled in the art should understand that the technical solutions of the present invention may be modified or replaced by equivalents without departing from the spirit and scope of the technical solutions of the present invention, which should all be included in the scope of the claims of the present invention.

Claims

1. A railway freight train condition assessment method based on time series and spectrum analysis, characterized by: include, Collect time series data and image data during train operation, and pre-process the time series data and image data; Stitch and register the preprocessed image data to generate full-scene images and extract dynamic visual features; For the preprocessed time series data, multi-scale time series spectrum joint analysis is performed, and multi-scale fusion is performed in combination with visual features to generate a multi-dimensional feature matrix; Build a health status baseline library, compare the multidimensional feature matrix with the health status baseline library, and identify abnormal features; Construct a train status assessment model, evaluate the train operation status according to abnormal characteristics, and generate a status assessment report based on the assessment results.

2. The railway freight train status assessment method based on time series and spectrum analysis as claimed in claim 1, characterized in that: The time series data includes vibration data, temperature data and load data during train operation; The image data includes a crack image, a wear image, and an oil stain image of the train appearance.

3. The railway freight train status assessment method based on time series and spectrum analysis as claimed in claim 2, characterized in that: The specific steps of preprocessing the time series data and image data are as follows: A bandpass filter is used to denoise the vibration data; The moving average method is used to smooth the temperature data; Use wavelet transform to remove random noise in drinking data; Image data is enhanced through histogram equalization; Image data is normalized by adjusting the image size and resolution.

4. The railway freight train status assessment method based on time series and spectrum analysis as claimed in claim 3 is characterized by: The pre-processed image data is stitched and registered to generate a full scene image and extract visual features. The specific steps are as follows: Use SIFT to identify feature points from preprocessed image data and extract feature descriptors from each feature point. Based on the feature descriptors, use the FLANN feature matching method to match feature point pairs between different image data. The RANSAC algorithm is used to randomly select the minimum number of feature point pairs from the matched feature point pairs, and a geometric transformation model is constructed based on the selected feature point pairs; According to the geometric transformation model, the position and angle of the image data are adjusted, the image data is spatially registered, and the registered image data is fused through a multi-resolution pyramid to form a full-scene image; Apply time series segmentation to the full scene image and use temporal optical flow to track the dynamic changes of the full scene image over time; Based on the dynamic changes of the full scene image over time, the dynamic visual features of the train are extracted from the full scene image using edge detection operators and CNN.

5. The railway freight train status assessment method based on time series and spectrum analysis as claimed in claim 4, characterized in that: The preprocessed time series data is subjected to multi-scale time series spectrum joint analysis, and multi-scale fusion is performed in combination with visual features to generate a multi-dimensional feature matrix. The specific steps are as follows: The vibration data is converted into time-frequency representation by short-time Fourier transform, and the dynamic spectrum features within a fixed time window are extracted based on the time-frequency representation; LSTM is used to capture the temporal dependencies between temperature data and load data, and to extract low-frequency trend features within a long time window. Through the multi-scale feature fusion method, the spectrum features, low-frequency trend features and visual features are fused to generate a multi-dimensional feature matrix, and the timestamp is added to the multi-dimensional feature matrix. The expression is: Among them, F is a multidimensional feature matrix, D N is the dynamic spectrum feature vector of the Nth sample point, L N is the low-frequency trend feature vector of the Nth sample point, V N is the dynamic visual feature vector of the Nth sample point, D1 is the dynamic spectrum feature vector of the first sample point, D2 is the dynamic spectrum feature vector of the second sample point, L1 is the low-frequency trend feature vector of the second sample point, L2 is the low-frequency trend feature vector of the second sample point, V1 is the dynamic visual feature vector of the first sample point, V2 is the dynamic visual feature vector of the second sample point, N is the total number of samples, t1 is the timestamp of the first sample point, t2 is the timestamp of the second sample point, t N is the timestamp of the Nth sample point, and T represents the transposition symbol.

6. The railway freight train status assessment method based on time series and spectrum analysis as claimed in claim 5, characterized in that: The health status baseline library is constructed, and the multi-dimensional feature matrix is ​​compared with the health status baseline library to identify abnormal features. The specific steps are as follows: Collect historical train operation data and conduct statistical analysis on the historical train operation data to identify the mean, variance and statistics of historical dynamic visual features, dynamic spectrum features and low-frequency trend features, and build a health status baseline library based on the identification results; The multidimensional feature matrix is ​​compared with the health status baseline library, and the outlier index of each feature in the multidimensional feature matrix is ​​predicted based on the interaction between the features. The expression is: Among them, A(F) is the outlier index of each sample point feature in the multidimensional feature matrix, D k is the dynamic spectrum feature vector in the kth time window, μ D is the mean vector of the dynamic spectrum feature, σ D is the standard deviation vector of the dynamic spectrum feature, L k is the low-frequency trend feature vector in the kth time window, μ L is the mean vector of the low-frequency trend feature, σ L is the standard deviation vector of low-frequency trend features, V k is the dynamic visual feature vector in the kth time window, μ V is the mean vector of dynamic visual features, σ V is the standard deviation vector of dynamic visual features, ω(t k ) is the time weighting factor, t k is the timestamp corresponding to the kth time window, λ is the weight coefficient of the interaction term, I(D k ,V k ) is the Pearson correlation coefficient between the dynamic spectrum feature vector and the dynamic visual feature vector, k is the index variable of the time window; Based on historical data, anomaly detection threshold E is defined. When A(F) is greater than E, the sample point features in the current time window are considered to be abnormal features.

7. The railway freight train status assessment method based on time series and spectrum analysis as claimed in claim 6, characterized in that: The Pearson correlation coefficient between the dynamic spectrum feature vector and the dynamic visual feature vector is expressed as: Among them, D k,i is the i-th element of the dynamic spectrum feature vector in the k-th time window, V k,i is the i-th element of the dynamic visual feature vector in the k-th time window, m represents the number of elements in each feature vector, and i represents the index variable of the element in each feature vector.

8. The railway freight train status assessment method based on time series and spectrum analysis as claimed in claim 6, characterized in that: The train status assessment model is constructed to assess the train running status according to the abnormal characteristics, and a status assessment report is generated according to the assessment results. The specific steps are as follows: Using LSTM as the basis, a multi-layer LSTM unit is set up. Each unit is responsible for capturing the local time dependency within a time window. The LSTM layer receives the abnormal features as input and outputs the time dependency representation of the abnormal features. The Transformer structure is combined with the LSTM layer, and the dynamic change characteristics in the time-dependent representation of abnormal features are processed through a feedforward neural network; The LSTM layer and the Transformer structure are integrated through the fully connected layer to form a train status assessment model; The abnormal features are input into the train status assessment model to predict the train operation status score. The expression is: Where S is the train running status score, Y is the score calibration coefficient, ΔB(t) is the abnormal feature vector at time point t, H(t) represents the dynamic weight of the train running status score at time point t, t0 represents the starting time of the time window, and t e represents the end time of the time window, ξ represents the weight coefficient of the abnormal feature vector modulus, ΔB k The abnormal feature vector in the kth time window, R k is the risk coefficient in the kth time window, λ is the weight coefficient of the train operation cost factor, C k is the train operation cost factor in the kth time window; Based on the statistical analysis of historical data, the health status benchmark value Q is defined; When S ≥ Q, the train is considered to be in good running condition; When S<Q, the train running status is considered abnormal; Based on the assessment results, a status assessment report is generated.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, wherein: When the processor executes the computer program, the steps of the railway freight train status assessment method based on time series and spectrum analysis described in any one of claims 1 to 8 are implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the steps of the railway freight train status assessment method based on time series and spectrum analysis described in any one of claims 1 to 8 are implemented.

Images